Fix article reference for Adrian Roselli (#4380)

* Fix article reference for Adrian Roselli

Adrian mentioned that he didn't explicitly say LLMs make image
descriptions better. This commit fixes that interpretation.

* Fix typo

fixes #4379

* Update src/content/en/2025/accessibility.md

---------

Co-authored-by: Barry Pollard <barrypollard@google.com>

Author: tricinel

src/content/en/2025/accessibility.md

@@ -858,7 +858,7 @@ Experts like Joe Dolson have explored whether <a hreflang="en" href="https://wpb
 
 Hidde de Vries <a hreflang="en" href="https://hidde.blog/ai-for-accessible-components/">contrasts how humans and language models approach accessible component code</a>. Humans base HTML, CSS, and ARIA decisions on specifications, user needs, assistive technology behavior, and platform quirks, all guided by intentions for the interface. LLMs instead predict likely code from training data, which is problematic because most existing code has accessibility issues, and the models lack intent or understanding of specific users.
 
-Adrian Roselli acknowledges that recent advances in computer vision and LLMs have brought real benefits, such as better image descriptions and improved captions and summaries. However, he argues <a hreflang="en" href="https://adrianroselli.com/2023/06/no-ai-will-not-fix-accessibility.html">these tools still lack context and authorship</a>. They can't know why content was created, what a joke or meme depends on, or how an interface is meant to work. Their descriptions and code suggestions can easily miss the point or mislead users.
+Adrian Roselli acknowledges that recent advances in computer vision and LLMs can potentially help readers distill complex articles into understandable summaries. However, he argues <a hreflang="en" href="https://adrianroselli.com/2023/06/no-ai-will-not-fix-accessibility.html">these tools still lack context and authorship</a>. They can't know why content was created, what a joke or meme depends on, or how an interface is meant to work. Their descriptions and code suggestions can easily miss the point or mislead users.
 
 AI raises significant ethical concerns that go beyond accessibility.
 
@@ -942,7 +942,7 @@ The map of TLD ranking is very similar to 2024, but obviously doesn't include th
 
 {{ figure_markup(
   image="map-accessible-countries-by-tld.png",
-  caption="Map of ccessible countries by Top Level Domain (TLD).",
+  caption="Map of accessible countries by Top Level Domain (TLD).",
   description="Displayed visually in a world map, the most accessible countries are Norway with 87%, Finland with 86%, followed by Canada, USA, UK, Sweden, Ireland, Australia, New Zealand, Austria, Belgium, Switzerland, Denmark, and South Africa. China is the least accessible by Top Level Domain, with close to 67%.",
   chart_url="https://docs.google.com/spreadsheets/d/e/2PACX-1vQFD-7C6Jv6q1JyviDsKosRlVwaok7g7nRCQ9NGMw5MaAAohL7EcDejVwgp13Z_T2S_57Zi0YaVb7st/pubchart?oid=1554186781&format=interactive",
   sheets_gid="1037208406",